Sora public-data review: the final rate card

What is being removed, and when

The wind-down ran in two stages and both have now passed. OpenAI's deprecations page gives an announcement date of 24 March 2026 and a shutdown date of 24 September 2026, and lists six entries under one heading.

That empty replacement column is the whole story for developers. A normal model retirement is a string swap — you change sora-2 to whatever OpenAI names next and redeploy. Here there is nothing to swap to, because the endpoint itself goes with the models. Migrating means adopting another vendor's API and rewriting the request shape and the job-polling loop around it, which is days of work rather than an afternoon.

The consumer half went first. The Sora app and website were discontinued on 26 April 2026, well under a year after the app launched, with users directed to export their library before the cutoff. If you never exported, assume that material is gone; there is no site left to log into. The shutdown was reported at the time by TechCrunch, dated the same day as the deprecation notice.

What OpenAI still charges, as of 21 September 2026

Sora has no plans, no credits and no free tier. It bills per second of generated video, and audio is produced with the picture at every rate. These figures are transcribed from OpenAI's model reference pages for sora-2 and sora-2-pro.

Two different things are being charged for, stacked. The first is the model: at an unchanged 720p, moving from sora-2 to sora-2-pro costs exactly three times as much, $0.10 to $0.30, for no extra pixels at all. The second is resolution, priced again on top — within Pro, 720p to 1024p is +66.7% and 1024p to 1080p a further +40%, so 720p to 1080p inside the same model is +133%. Multiply the two and the cheapest Sora second and the dearest differ by exactly 7x.

The practical consequence is that sora-2 has no upgrade path that only buys resolution. Standard is a 720p-only model. Wanting 1080p means buying Pro's quality premium whether you wanted it or not, and the bill goes up sevenfold.

Correction, 23 September 2026: there was a batch tier, at half price

The table above is the rate card this review published from 15 September, and it is the standard column only. OpenAI's API pricing page also carries a batch column for both video models, at exactly 50% of the standard rate on every configuration, and CartSignal had not published it. The batch route is documented in OpenAI's video generation guide, which instructs callers to "Use the Batch API when you need to queue many video renders for offline processing, review pipelines, or studio workflows" and notes that "Batch-generated videos are available for download for up to 24 hours after the batch completes" — a longer window than the one-hour URLs standard generations get.

What this changes. The claim below that Sora "has no cheap tier" is true of the standard column and false of the card as a whole: on batch its floor is $0.05 a second — level with Veo 3.1 Lite rather than double it, so the model the market spent this month leaving held a joint-cheapest second in the field. The documented worst case halves too: a 20-second sora-2-pro generation at 1080p was $14.00 standard and $7.00 on batch, and the 120-second extended maximum $84.00 and $42.00. Every exit from Sora is still cheaper or break-even measured from the standard rates, but measured from the batch rates those margins roughly halve, and from the batch floor the best available exit is a tie rather than a saving.

Where it was hiding extends this page's own theme rather than contradicting it. The batch column is on the pricing page. It is not on the sora-2 or sora-2-pro model reference pages, which were re-read today and show a single standard column — the same two pages that, one day from removal, still carry no deprecation notice. That is now the third distinct thing OpenAI's video surfaces disagree about, after the seconds range and the size list. Whether Google matches the discount is unresolved: its Batch API documentation states batch usage "is priced at 50% of the standard interactive API cost for the equivalent model" but lists no models, and the Gemini API pricing page shows no batch column for any Veo row. Full field in what AI video generation costs per second.

How long a clip can actually be, and why this review got it wrong

Corrected 21 September 2026. This section previously reported a hard twelve-second ceiling and concluded that no single generation could cost more than $8.40. That was sourced from one OpenAI page. Reading three of them side by side shows they do not agree, and the one this review relied on is the outlier.

Two of the three agree with each other and disagree with the API reference. Taking the majority reading, the ceiling on a single request is 20 seconds, not 12, and a video can be extended six times to 120 seconds total. That moves two numbers this site published:

The conclusion that went with the old number has to go too. This review said Sora "was always a shot generator, not a scene generator", and set its 12 seconds against Seedance 2.5's documented 30 seconds in a single pass. On the documented behaviour that is backwards at the length that matters: 120 seconds is 4x Seedance's single-pass 30, 8x Kling's 15-second maximum and 15x the 8 seconds Veo bills per generation in Flow. What survives is the narrower point — in one uninterrupted pass Seedance's 30 seconds is still 1.5x Sora's 20, and an extended Sora video is six joins rather than one take.

The honest caveat, and it is a real one. Documented is not the same as available. A thread on OpenAI's own developer forum dated 16 March 2026 reports the API returning a 400 error calling 16 seconds "an invalid integer when extending a video", with a community responder noting the longer durations "are not supported in the API libraries, and are not present in the video playground". No OpenAI staff reply appears in that thread and CartSignal has run no requests, so which ceiling was actually enforced cannot be settled from public data. With removal due on 24 September 2026 it is not worth finding out experimentally: any new work should be priced against another vendor.

Sora against the field, per second

Google publishes per-second rates for Veo 3.1 on the Gemini API pricing page, re-verified today, with audio included by default and no free tier. That makes it the one direct like-for-like on billing units, and the comparison is unflattering for OpenAI at every point except one.

Three things fall out of one sorted list. Sora has no cheap tier on this column: its standard floor is $0.10 a second where Google's is $0.05, so the least you can spend on a standard Sora second is exactly double the least you can spend on a Veo one — but see the batch correction above, which brings Sora's real floor level with Google's at $0.05. They match on price at the same resolution exactly once, at $0.10 for a 720p second with audio, where sora-2 and Veo 3.1 Fast are identical to the cent. And the top of the card belongs to OpenAI: $0.70 for a 1080p second is the most expensive rate on this page, 75% above Veo 3.1 Standard at the same resolution and 5.8 times Veo 3.1 Fast's $0.12. Google's single dearest second of video — Standard at 4K, $0.60 — still comes in 14.3% below OpenAI's 1080p one. The sorted list throws up one more collision worth noticing: sora-2-pro at 720p and Veo 3.1 Fast at 4K are both $0.30 a second — the same money buys 720p from one vendor and 4K from the other.

A note on what that does and does not say. These are list prices for a second of output, not a quality ranking; CartSignal has generated no video and runs no prompt-adherence or motion tests. Two models at the same price may be nothing like each other on screen. What the table establishes is narrower and still useful: at no point on its own rate card did Sora hold a price advantage over the incumbent it was being compared against. Both these figures and Veo's sit inside the wider generative lane costed on what AI video generation costs per second, which runs from $3.00 to $90.00 a finished minute across Veo, Kling, Runway and Seedance as well.

Every exit is cheaper

Forced migrations usually cost money. This one does not, at least on the meter. Taking each Sora row and the nearest Veo row at the same resolution:

The worst case is break-even. There is no row where a buyer pays more to leave. Whatever else the migration costs you in engineering time — and the empty replacement column means it is real engineering time — the recurring bill does not go up, and for anyone who was on Pro it falls sharply. That is the reverse of the usual shutdown, where the cheap incumbent dies and its users discover what the market rate is.

What the documentation said at the end

All of this was re-checked on 24 September 2026, the removal date, and is published as observation rather than as a ruling on which OpenAI page was right. Every presence and absence below was confirmed by searching the literal page source, for the reason given in the first bullet.

• The warning existed, and this review previously said it did not. The video generation guide opens with a banner reading "The Sora 2 video generation models and Videos API are deprecated and will shut down on September 24, 2026," naming all five model strings and linking the deprecations page. The Internet Archive holds a copy carrying that banner on 11 September 2026. This page nonetheless claimed on 15, 21 and 23 September that the guide carried no notice, because the tool used to read it summarises pages and silently drops callouts — asked again today, it still answered that no notice appears. That is a methodology failure rather than a fact that changed, it is corrected here, and the standard on this site is now that any claim about something being absent from a vendor page is verified against raw source.
• The corrected count is three surfaces without a notice, and they are the ones with the prices on. On removal day the sora-2 and sora-2-pro model references both showed a full rate card and live per-tier rate limits — 25 requests per minute on Tier 1 rising to 375 on Tier 5 — with no shutdown notice; the only deprecation language anywhere in either page was a row in the snapshot table marking the older sora-2-2025-12-08 snapshot as superseded, which is a different thing entirely. The API pricing page carried the whole video table, standard and batch, with no notice. The prompting cookbook contains no deprecation language at all. So the guide told you, and the three pages a buyer or developer actually reads for numbers did not.
• The 1080p puzzle is now resolvable, and it was the reference page. This review previously noted that sora-2-pro is priced at $0.70 a second for 1080x1920 and 1920x1080 while the Videos API reference omits that pair from the allowed values of size, and said CartSignal could not establish which page was stale. It can now: OpenAI's own prompting cookbook independently lists sora-2-pro as supporting 720x1280, 1280x720, 1024x1792, 1792x1024, 1080x1920 and 1920x1080, corroborating the model page's rate card. Two sources against one, and the outlier is the same page that carries the stale seconds list.
• That is the pattern worth taking away. Both discrepancies on this page trace to a single surface — the API reference for the create endpoint — and it is the worst one to have wrong, because it is the page a developer reads while writing the request. On this evidence it lagged the product on seconds and on size simultaneously, and was never brought back into line before the withdrawal.
• What happens to a request after the 24th is still not documented. The deprecations page says these entries are removed from the API on that date and says nothing about what a call made afterwards returns — no status code, no error shape, no grace period. Searching the page source for "410" returns nothing. So there is no documented failure to code a graceful fallback against, which matters more than it sounds: see below.

OpenAI's video generation guide separately notes that the remix endpoint, previously used to edit an existing generation, "is being deprecated" in favour of the edits endpoint — so the editing path was already moving before the models went.

What a call to the dead endpoint returns now

OpenAI has not published the answer, and the confident answer circulating online is not sourced to anything OpenAI wrote.

Several migration guides published in the days around the shutdown state a specific behaviour: that from 25 September every request to a /videos/ URL returns 410 Gone. One of them attributes that to OpenAI's deprecations page. CartSignal searched the source of that page on 24 September 2026 and it contains no "410", no HTTP status code of any kind, and no description of post-removal behaviour — only the removal date and the empty replacement column. The claim may still turn out to be right, but it is not something OpenAI documented, and it has been repeated between sites without anyone checking the citation.

The only published evidence points the other way. When OpenAI has retired model strings before, developers report HTTP 404 with model_not_found — that is the documented error code for a model identifier the API does not recognise, and it is what the developer forum filled with when earlier snapshots were withdrawn. The Videos API is a slightly different case because an entire endpoint is being removed rather than a model alias, and a removed route plausibly returns 404 for a different reason; during an unrelated outage in April 2026 the endpoint returned 404 Invalid URL. Either way, 404 is the behaviour with precedent behind it and 410 is the behaviour with a blog post behind it.

The practical instruction is the same whichever it is, and it is the reason this matters: do not write a handler that branches on a status code you read in a migration guide. Code the failure as "no video came back" — catch the request failing, log the body, and fall through to your replacement vendor or a queue. A fallback keyed to 410 that meets a 404 fails exactly as badly as no fallback at all.

A second claim travelling with the first is that OpenAI permanently deletes all stored Sora generations today. That one carries no citation anywhere CartSignal could find, and it cannot be checked here because help.openai.com returns HTTP 403 to our fetches, with a browser user agent as well as without — consistent with every previous attempt recorded on this site. It is also close to moot for API users, for the reason in the next section: there was nothing stored to delete.

What happens to video you already generated

This is the question with a deadline attached, and for API users the answer is more urgent than the shutdown date — because the relevant window is measured in hours and always was.

OpenAI's video generation guide states plainly: "Download URLs are valid for a maximum of 1 hour after generation. If you need long-term storage, copy the file to your own storage system promptly." For work submitted through the Batch API the guide gives a longer window — "Batch-generated videos are available for download for up to 24 hours after the batch completes". The API reference adds a GET /videos endpoint returning "recently generated videos for the current project", and an expires_at field described as the "Unix timestamp (seconds) for when the downloadable assets expire, if set".

The practical reading: there is no archive on OpenAI's side for the shutdown to take away. If your pipeline copied each MP4 to your own storage at generation time, as the guide instructs, nothing was at risk today and you lost only the ability to make new video. If it did not — if your code kept OpenAI URLs and re-fetched them on demand — those links were already expiring an hour after each generation, long before any of this was announced, so the breakage predates the shutdown and outlives it. This is also why the deletion claim in the previous section changes little for API callers: the one-hour window had already done the deleting.

The consumer side is a separate matter and CartSignal cannot verify it first-party: help.openai.com returned HTTP 403 to our fetch again today, consistent with this site's previous attempts. Third-party guides report an export tool for app and web generations, reached from a sunset page on the Sora domain, and report that it covers material created in the app rather than through the API. Both of those are reported rather than confirmed here. They are also moot for anyone reading this page for API reasons, and the app itself closed on 26 April 2026.

If Sora was in your stack

• The date was 24 September 2026, not "September". Requests to the Videos API and every sora-2 alias stopped being served on that day. Anything scripted against it — a no-code automation, a scheduled job, a product feature — simply stops returning video unless you built a fallback, and OpenAI never published what the failure looks like, so do not key that fallback to a status code.
• Check where your finished files live before you do anything else. Download URLs expired an hour after generation. If your code stored OpenAI URLs rather than copies, that is a problem that was already happening and is not fixed by migrating.
• Aggregators are not a reprieve. Resellers that proxy the official API lose access exactly when you do, because it is the same endpoint behind them. Treat any offer of continued sora-2 access after the removal date as unverified, and do not build on it.
• Price the replacement on your own resolution, not the headline. If you were on 720p standard, Veo 3.1 Fast is a straight swap on cost and Lite halves it. If you were on Pro at 1080p, almost anything is cheaper. Full breakdown of both routes into Veo — the per-second API and the credit-based Flow studio — is on the Google Veo profile.
• Re-check your clip length assumptions — in both directions. Vendors differ more on maximum take length than they differ on price, and as the correction above shows, the documented Sora ceiling is longer than this site previously reported. If your pipeline stitches 12-second segments, a destination allowing longer single takes may make that logic redundant; if it relied on extending to a minute or more, most alternatives cap a single generation far shorter. Kling and Runway are the other two most commonly named destinations; Runway resells several third-party models on one credit meter, including Seedance.
• Do not forget the labelling duty. Changing generator does not change your obligations. Since 2 August 2026, EU AI Act Article 50 requires generative output to be marked machine-readably and requires you to disclose synthetic media you publish — see labelling AI video and voice under the EU AI Act.

The full migration shortlist, with each destination's plan prices and credit mechanics, is on Sora alternatives.

FAQ

Is the Sora API still available?

No. 24 September 2026 is the removal date OpenAI set for the Videos API and every sora-2 alias, and the consumer app and website closed earlier, on 26 April 2026. OpenAI's deprecations page, announced 24 March 2026, lists the Videos API and the strings sora-2, sora-2-pro, sora-2-2025-10-06, sora-2-2025-12-08 and sora-2-pro-2025-10-06 for removal, with a dash in the recommended-replacement column for every one. OpenAI now sells no video generation model at all, so there is no model string to switch to and migration means changing vendor.

Does the Sora API return a 410 error now?

OpenAI has not documented it, and the widely repeated "410 Gone" answer is not sourced to OpenAI. Several migration guides published around the shutdown state that requests to /videos/ return 410, and at least one attributes that to OpenAI's deprecations page — but that page, checked in its literal source on 24 September 2026, contains no "410" and no description of post-removal behaviour at all. The only pattern with evidence behind it runs the other way: when OpenAI has withdrawn model strings before, developers report HTTP 404 with model_not_found. Whichever it turns out to be, do not write a fallback that branches on the status code. Treat the response as "no video came back" and fall through to your replacement vendor, because a handler keyed to 410 that meets a 404 fails as completely as having no handler.

How much does the Sora API cost per second?

sora-2 is $0.10 a second and exists only at 720x1280 and 1280x720. sora-2-pro is $0.30 at those sizes, $0.50 at 1024x1792 and 1792x1024, and $0.70 at 1080x1920 and 1920x1080. Audio is included at every rate. As finished minutes that is $6.00, $18.00, $30.00 and $42.00. Added 23 September 2026: OpenAI's pricing page also publishes a batch tier at exactly half each of those rates — $0.05, $0.15, $0.25 and $0.35, or $3.00 to $21.00 a finished minute — on a column that appears on the pricing page but on neither model reference page. Model and resolution are priced separately — Pro costs exactly 3x standard at an unchanged 720p — so the card's floor and ceiling differ by exactly 7x on either column.

Will migrating off Sora cost more?

No. Every exit is cheaper or the same on standard rates. Veo 3.1 Fast at 720p is $0.10, identical to sora-2's only rate; Veo 3.1 Lite at $0.05 halves it. Qualified 23 September 2026: from the batch tier at $0.05 the margins roughly halve and the best exit is a tie, since Veo 3.1 Lite matches it and Google publishes no batch discount for Veo. At the top, sora-2-pro's $0.70 second at 1080p is 75% above Veo 3.1 Standard at the same resolution and 5.8 times Veo 3.1 Fast. Google's most expensive second of all, Standard at 4K for $0.60, is still 14.3% below OpenAI's 1080p second. These are prices, not quality judgements — no video was generated for this review.

How long can a single Sora clip be?

OpenAI's own pages disagree. The Videos API reference gives seconds allowed values of 4, 8 and 12, but the video generation guide says both models "support 16- and 20-second generations" and allows six extensions to "a maximum total length of 120 seconds", and the prompting cookbook lists 4, 8, 12, 16 and 20. Taking the two agreeing sources, a single generation runs to 20 seconds and costs at most $14.00 at Pro 1080p, and an extended 120-second video costs up to $84.00. A developer forum thread from March 2026 reports 16 seconds being rejected with a 400 error, so documented and available may differ, and CartSignal has run no requests.

Do I lose videos I already made when the Sora API shuts down?

Not on the 24th, because there is nothing there to lose. OpenAI's video generation guide states that "Download URLs are valid for a maximum of 1 hour after generation" and tells callers to copy files to their own storage promptly; batch output is available for up to 24 hours. So API output was never held in a durable library — if your pipeline saved each MP4 at generation time you keep everything, and if it stored OpenAI URLs instead those were already expiring hourly. The separate consumer export route is reported by third parties and could not be verified here, as help.openai.com returns HTTP 403 to our fetches.

Has CartSignal tested Sora?

No. This is a public-data review built from OpenAI's deprecations page, the sora-2 and sora-2-pro model reference pages, the video generation guide, the Videos API reference and the Sora 2 prompting cookbook, all re-checked 21 September 2026, plus contemporaneous reporting of the March 2026 announcement and one dated OpenAI developer forum thread. No video was generated and no quality claim is verified. Per-minute conversions, percentages and cross-vendor comparisons are CartSignal arithmetic over each vendor's own published rates.

Source links

• Official OpenAI deprecations page (announcement 24 March 2026, shutdown 24 September 2026, all six entries, empty replacement column)
• Official sora-2 model reference ($0.10/second, 720p sizes, synced audio, per-tier rate limits)
• Official sora-2-pro model reference ($0.30 / $0.50 / $0.70 per second by resolution)
• Official Videos API reference (seconds allowed values 4, 8, 12; size allowed values; model options) — the surface this review identifies as stale on both seconds and size
• Official OpenAI API pricing page (the video table: standard $0.10 / $0.30 / $0.50 / $0.70 and batch $0.05 / $0.15 / $0.25 / $0.35 per second, still live and unmarked on the removal date)
• Internet Archive capture of the video generation guide, 11 September 2026 — evidence that the shutdown banner predates this review saying it was absent
• Official video generation guide (the deprecation banner naming all five model strings, 16- and 20-second generations, extension to 120 seconds, the one-hour download URL window, batch 24-hour window, remix endpoint deprecation)
• Official Sora 2 prompting guide, OpenAI cookbook (seconds 4/8/12/16/20; sora-2-pro sizes including the 1080p pair)
• Official Videos API reference index (GET /videos list endpoint, /videos/{id}/content download, expires_at field)
• OpenAI developer forum thread, 16 March 2026 — 400 error on a 16-second extension, no staff reply (community report, not an OpenAI statement)
• Official Gemini API pricing (Veo 3.1 Standard, Fast and Lite per-second rates, audio included, no free tier)
• TechCrunch, 24 March 2026 — contemporaneous reporting of the shutdown announcement